pith. sign in

arxiv: 2606.30310 · v1 · pith:DI3TXOGNnew · submitted 2026-06-29 · 📊 stat.ML · cs.LG

Highly Data Parallelizable Estimation of the Sliced-Wasserstein Distance Using Cumulative Distribution Functions

Christophe Vauthier , Quentin M\'erigot , Anna Korba This is my paper

Pith reviewed 2026-06-30 04:02 UTC · model grok-4.3

classification 📊 stat.ML cs.LG
keywords sliced wasserstein distancecumulative distribution functionsmonte carlo estimationparallel computationfederated learningoptimal transport
0
0 comments X

The pith

Sliced Wasserstein distance estimators built from cumulative distribution functions of projections avoid sorting and enable massive parallelism.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a class of estimators for the sliced Wasserstein distance that rely on cumulative distribution functions of one-dimensional projections rather than quantile functions obtained by sorting. These estimators compute the distance through averages of one-dimensional transport costs derived from the CDFs, with some variants controlled by hyperparameters for variance or smoothness. The approach is shown to suit cases where CDFs are easier to handle than quantiles, such as Gaussian mixtures, and to fit federated settings because local CDFs can be aggregated without moving raw samples. A sympathetic reader would care if the method removes a key computational bottleneck that has limited sliced Wasserstein use on very large or distributed data.

Core claim

A new class of estimators for the Sliced Wasserstein distance is constructed from cumulative distribution functions of projected measures; these estimators replace the standard Monte Carlo procedure that sorts samples to obtain quantiles, thereby eliminating the sorting step and supporting computation across disjoint data shards without exchanging the original points.

What carries the argument

CDF-based estimators for the sliced Wasserstein distance, which replace quantile functions with cumulative distribution functions of projected measures to compute one-dimensional Wasserstein distances.

Load-bearing premise

The statistical bias, variance, and consistency of the CDF-based estimators are comparable to or better than those of the conventional quantile-based Monte Carlo estimator.

What would settle it

Run both the standard sorted-quantile estimator and each CDF-based estimator on the same collection of random projections of a fixed dataset and compare their empirical bias and variance to the true sliced Wasserstein value.

Figures

Figures reproduced from arXiv: 2606.30310 by Anna Korba, Christophe Vauthier, Quentin M\'erigot.

Figure 1
Figure 1. Figure 1: Plots of (Fµ(x) − Fν(y))+ on the tri￾angle 0 ≤ x ≤ y ≤ 1 (bottom) and of the probability density functions and CDFs of µ, ν (top) for various pairs (µ, ν). The measures µ and ν are obtained by truncating normal distribu￾tions N (m−, σ) and N (m+, σ) to [−1, 1]. Left: m± = ±0.2 and σ = 1. Middle: m± = ±0.9 and σ = 1. Right: m± = ±0.9 and σ = 0.1. Variance minimization via importance sampling One drawback of… view at source ↗
Figure 2
Figure 2. Figure 2: Left: Results of fitting a Gaussian mixture model with 10 modes to a "ring-line-square" for [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of the values of SWd2 2,L,ηk for various choices of k. The red lines denotes the mean of all the computations of the estimators, which is expected to be close to the true SW2 2 value. However, this results does not tell us the concrete value of k that we should choose in practice. To guide our choice, we perform the following simple experiment. We sample two clouds of N = 1000 points in the pl… view at source ↗
Figure 4
Figure 4. Figure 4: Accuracy vs computing time for the different [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Gradient descent on the smoothed CDF-based estimators [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Samples from the GANs trained on MNIST • The CDF-based estimator SWd2 2,L,ε with L = 100000 and ε = 0.01. • The CDF-based estimator SWd2 2,L,ηk,ε with L = 100000, k = 0.05 and ε = 0.01. FID scores We compute FID scores using the pytorch-fid Python package. Each FID score is computed based on 10000 random samples from the trained model and the full training dataset. Full results [PITH_FULL_IMAGE:figures/fu… view at source ↗
Figure 7
Figure 7. Figure 7: Samples from the GANs trained on CelebA D.1 The OTDD and MMDSW dataset distances The Optimal Transport Dataset Distance (OTDD, [Alvarez-Melis and Fusi, 2021]) is a distance between labeled datasets of the form D = {(xi , li)} N i=1 ⊆ R d × {1, . . . , L}. It is defined the following way: for every label l ∈ {1, . . . , L}, let νD,l := 1 #{i, xi = l} X xi=l δxi ∈ P(R d ) (92) be the empirical measure of the… view at source ↗
Figure 8
Figure 8. Figure 8: Distance matrices for MMDSW with MNIST (top row) and CIFAR-10 (bottom row), with [PITH_FULL_IMAGE:figures/full_fig_p032_8.png] view at source ↗
read the original abstract

The Sliced Wasserstein (SW) distance has emerged as a computationally attractive alternative to the Wasserstein distance by leveraging one-dimensional optimal transport along random projections. Standard estimators of the SW distance rely on Monte Carlo averages of one-dimensional Wasserstein distances computed via quantile functions, which require sorting projected samples and access to full datasets. In this work, we introduce a new class of estimators for the Sliced Wasserstein distance based on cumulative distribution functions (CDFs) of projected measures, that avoid sorting and scale via massive dataset parallelism. This class includes several estimators, some of them being indexed by hyperparameters controlling their variance or smoothness. We show that they are especially well suited to scenarios in which CDFs are more tractable than quantile functions, such as mixtures of Gaussians, and moreover that they are also naturally compatible with federated learning, since CDFs of projected data can be computed and aggregated locally without requiring the exchange of raw samples.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper introduces a new class of estimators for the Sliced Wasserstein distance based on cumulative distribution functions (CDFs) of projected measures. These estimators are claimed to avoid sorting, enable massive dataset parallelism, and be especially suitable for mixtures of Gaussians and federated learning settings where CDFs can be computed and aggregated locally without exchanging raw samples.

Significance. If the statistical properties (consistency, bias, variance) of the CDF-based estimators can be established as comparable or superior to standard quantile-based Monte Carlo estimators, the approach would provide a meaningful advance for scalable and distributed computation of sliced Wasserstein distances in large-scale and privacy-sensitive machine learning applications.

major comments (1)
  1. The abstract and manuscript description introduce the CDF-based estimators and assert their suitability as replacements for quantile-based Monte Carlo SW estimators, but supply no derivation, consistency proof, convergence rates, bias/variance bounds, or empirical comparison. This analysis is load-bearing for the central claim that the new estimators are statistically reliable.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We agree that a rigorous statistical characterization is essential to support the claim that the CDF-based estimators can serve as reliable replacements for quantile-based Monte Carlo estimators, and we will supply the missing analysis in revision.

read point-by-point responses

  1. Referee: The abstract and manuscript description introduce the CDF-based estimators and assert their suitability as replacements for quantile-based Monte Carlo SW estimators, but supply no derivation, consistency proof, convergence rates, bias/variance bounds, or empirical comparison. This analysis is load-bearing for the central claim that the new estimators are statistically reliable.

    Authors: We agree with the referee that the current manuscript lacks the requested theoretical derivations, consistency proofs, convergence rates, bias/variance bounds, and direct empirical comparisons. These elements were not included in the initial submission. In the revised version we will add (i) a consistency proof and convergence-rate analysis for the CDF-based estimators, (ii) explicit bias and variance bounds (including the effect of the smoothing hyper-parameter), and (iii) a set of controlled experiments that compare the new estimators against the standard quantile-based Monte Carlo estimator on both synthetic and real data. This will directly substantiate the central claim. revision: yes

Circularity Check

0 steps flagged

No circularity; new CDF-based estimators introduced as independent construction

full rationale

The paper defines a new class of SW estimators via CDFs of projections (avoiding quantiles/sorting) and notes their suitability for Gaussians and federated settings. No load-bearing step reduces a claimed property (consistency, variance, parallelism) to a fitted input, self-citation, or definitional tautology. The absence of bias/variance derivations is a proof gap, not circularity. The chain is a direct proposal with external applicability claims.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no free parameters, axioms, or invented entities are stated or implied beyond standard assumptions of optimal transport and random projections.

pith-pipeline@v0.9.1-grok · 5697 in / 1012 out tokens · 41031 ms · 2026-06-30T04:02:00.162463+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 4 canonical work pages

  1. [1]

    Advances in Neural Information Processing Systems , volume=

    Non-asymptotic convergence bounds for Wasserstein approximation using point clouds , author=. Advances in Neural Information Processing Systems , volume=

  2. [2]

    arXiv preprint arXiv:2301.11235 , year=

    Handbook of. 2023 , keywords =. doi:10.48550/arXiv.2301.11235 , author =

    work page doi:10.48550/arxiv.2301.11235 2023

  3. [3]

    Abel Integral Equations , doi =

    Rudolf Gorenflo, Sergio Vessella , month = feb, year =. Abel Integral Equations , doi =

  4. [4]

    Acta Arithmetica , author =

    Low-discrepancy point sets for non-uniform measures , doi =. Acta Arithmetica , author =. 2013 , note =

    2013

  5. [5]

    2025 , eprint=

    Properties of Wasserstein Gradient Flows for the Sliced-Wasserstein Distance , author=. 2025 , eprint=

    2025

  6. [6]

    One-dimensional empirical measures, order statistics, and

    Bobkov, Sergey and Ledoux, Michel , month = sep, year =. One-dimensional empirical measures, order statistics, and. doi:10.1090/memo/1259 , note =

    work page doi:10.1090/memo/1259

  7. [7]

    2015 , publisher=

    Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling , author=. 2015 , publisher=

    2015

  8. [8]

    2012 , doi =

    Burkard, Rainer and Dell'Amico, Mauro and Martello, Silvano , title =. 2012 , doi =

    2012

  9. [9]

    2021 , eprint=

    Advances and Open Problems in Federated Learning , author=. 2021 , eprint=

    2021

  10. [10]

    The Twelfth International Conference on Learning Representations , year=

    Federated Wasserstein Distance , author=. The Twelfth International Conference on Learning Representations , year=

  11. [11]

    2025 , eprint=

    Federated Sinkhorn , author=. 2025 , eprint=

    2025

  12. [12]

    Journal of Mathematical Imaging and Vision , volume=

    Sliced and radon wasserstein barycenters of measures , author=. Journal of Mathematical Imaging and Vision , volume=. 2015 , publisher=

    2015

  13. [13]

    International Conference on Learning Representations , year=

    Sliced Wasserstein auto-encoders , author=. International Conference on Learning Representations , year=

  14. [14]

    International Conference on Machine Learning , pages=

    Sliced-Wasserstein flows: Nonparametric generative modeling via optimal transport and diffusions , author=. International Conference on Machine Learning , pages=. 2019 , organization=

    2019

  15. [15]

    Scale Space and Variational Methods in Computer Vision: Third International Conference, SSVM 2011, Ein-Gedi, Israel, May 29--June 2, 2011, Revised Selected Papers 3 , pages=

    Wasserstein barycenter and its application to texture mixing , author=. Scale Space and Variational Methods in Computer Vision: Third International Conference, SSVM 2011, Ein-Gedi, Israel, May 29--June 2, 2011, Revised Selected Papers 3 , pages=. 2012 , organization=

    2011

  16. [16]

    Advances in Neural Information Processing Systems , volume=

    Statistical and topological properties of sliced probability divergences , author=. Advances in Neural Information Processing Systems , volume=

  17. [17]

    Electronic Communications in Probability , year=

    Strong equivalence between metrics of Wasserstein type , author=. Electronic Communications in Probability , year=

  18. [18]

    Advances in Neural Information Processing Systems , volume=

    Sliced mutual information: A scalable measure of statistical dependence , author=. Advances in Neural Information Processing Systems , volume=

  19. [19]

    Advances in Neural Information Processing Systems , volume=

    Statistical, robustness, and computational guarantees for sliced Wasserstein distances , author=. Advances in Neural Information Processing Systems , volume=

  20. [20]

    Advances in neural information processing systems , volume=

    Generalized sliced wasserstein distances , author=. Advances in neural information processing systems , volume=

  21. [21]

    and Hoffmann, Heiko , title =

    Kolouri, Soheil and Rohde, Gustavo K. and Hoffmann, Heiko , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

  22. [22]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Max-sliced wasserstein distance and its use for gans , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  23. [23]

    Advances in Neural Information Processing Systems , volume=

    Energy-based sliced Wasserstein distance , author=. Advances in Neural Information Processing Systems , volume=

  24. [24]

    International conference on machine learning , pages=

    Alvarez-Melis, David and Fusi, Nicol. International conference on machine learning , pages=. 2021 , organization=

    2021

  25. [25]

    Asian Conference on Machine Learning , pages=

    Sliced Wasserstein variational inference , author=. Asian Conference on Machine Learning , pages=. 2023 , organization=

    2023

  26. [26]

    International Conference on Learning Representations , year=

    Distributional Sliced-Wasserstein and Applications to Generative Modeling , author=. International Conference on Learning Representations , year=

  27. [27]

    International Conference on Learning Representations , year=

    Hierarchical sliced wasserstein distance , author=. International Conference on Learning Representations , year=

  28. [28]

    International Conference on Machine Learning , pages=

    Shedding a PAC-Bayesian light on adaptive sliced-Wasserstein distances , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  29. [29]

    Advances in Neural Information Processing Systems , volume=

    Markovian sliced Wasserstein distances: Beyond independent projections , author=. Advances in Neural Information Processing Systems , volume=

  30. [30]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Sliced wasserstein discrepancy for unsupervised domain adaptation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  31. [31]

    ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Approximate Bayesian computation with the sliced-Wasserstein distance , author=. ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2020 , organization=

    2020

  32. [32]

    The 22nd International Conference on Artificial Intelligence and Statistics , pages=

    Orthogonal estimation of Wasserstein distances , author=. The 22nd International Conference on Artificial Intelligence and Statistics , pages=. 2019 , organization=

    2019

  33. [33]

    International Conference on Machine Learning , year=

    Sliced iterative normalizing flows , author=. International Conference on Machine Learning , year=

  34. [34]

    International Conference on Learning Representations , year=

    Distance-based tree-sliced Wasserstein distance , author=. International Conference on Learning Representations , year=

  35. [35]

    Advances in neural information processing systems , volume=

    Tree-sliced variants of Wasserstein distances , author=. Advances in neural information processing systems , volume=

  36. [36]

    The Thirteenth International Conference on Learning Representations , year=

    Tree-Wasserstein Distance for High Dimensional Data with a Latent Feature Hierarchy , author=. The Thirteenth International Conference on Learning Representations , year=

  37. [37]

    Statistics and Computing , author =

    A tutorial on spectral clustering , volume =. Statistics and Computing , author =. 2007 , pages =. doi:10.1007/s11222-007-9033-z , abstract =

    work page doi:10.1007/s11222-007-9033-z 2007

  38. [38]

    2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Generative Modeling Using the Sliced Wasserstein Distance , author=. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    2018

  39. [39]

    International Conference on Learning Representations , year=

    Spectral Normalization for Generative Adversarial Networks , author=. International Conference on Learning Representations , year=

  40. [40]

    Proceedings of the 36th International Conference on Neural Information Processing Systems , articleno =

    Nguyen, Khai and Ho, Nhat , title =. Proceedings of the 36th International Conference on Neural Information Processing Systems , articleno =. 2022 , isbn =

    2022

  41. [41]

    Bonnotte, Nicolas , year=

  42. [42]

    2016 ,booktitle =

    Practical Secure Aggregation for Federated Learning on User-Held Data ,author =. 2016 ,booktitle =

    2016

  43. [43]

    Differential Privacy: A Survey of Results

    Dwork, Cynthia. Differential Privacy: A Survey of Results. Theory and Applications of Models of Computation. 2008

    2008

  44. [44]

    Dwork, Cynthia and Roth, Aaron , title =. Found. Trends Theor. Comput. Sci. , month = aug, pages =. 2014 , issue_date =. doi:10.1561/0400000042 , abstract =

    work page doi:10.1561/0400000042 2014

  45. [45]

    Kingma and Jimmy Ba , editor =

    Diederik P. Kingma and Jimmy Ba , editor =. Adam:. 3rd International Conference on Learning Representations,. 2015 , timestamp =

    2015

  46. [46]

    LeCun, Yann and Cortes, Corinna , groups =

  47. [47]

    Deep Learning Face Attributes in the Wild , year=

    Liu, Ziwei and Luo, Ping and Wang, Xiaogang and Tang, Xiaoou , booktitle=. Deep Learning Face Attributes in the Wild , year=

  48. [48]

    Learning multiple layers of features from tiny images , year =

    Krizhevsky, Alex and Hinton, Geoffrey , address =. Learning multiple layers of features from tiny images , year =

  49. [49]

    2011 , MONTH = Oct, KEYWORDS =

    Pedregosa, Fabian and Varoquaux, Ga. 2011 , MONTH = Oct, KEYWORDS =

    2011

  50. [50]

    ArXiv , year=

    Federated Learning with Personalization Layers , author=. ArXiv , year=

  51. [51]

    Proceedings of the 38th International Conference on Machine Learning , pages =

    Exploiting Shared Representations for Personalized Federated Learning , author =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , editor =

    2021

  52. [52]

    Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =

    Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp , title =. Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =. 2017 , isbn =

    2017

  53. [53]

    Forty-second International Conference on Machine Learning , year=

    Flowing Datasets with Wasserstein over Wasserstein Gradient Flows , author=. Forty-second International Conference on Machine Learning , year=

  54. [54]

    Transactions on Machine Learning Research , issn=

    Slicing Unbalanced Optimal Transport , author=. Transactions on Machine Learning Research , issn=. 2025 , note=

    2025